Agency Runtime v10

Purpose

Turn completed v9 executions into deterministic benchmark observations that influence the next plan without weakening host security or promotion controls.

Feedback loop

Use:
PROFILE -> ADAPT FROM HISTORY -> ROUTE -> RESOURCE PLAN -> EXECUTE (HOST) -> EVIDENCE -> VERIFY -> REGRESSION -> RELEASE -> BENCHMARK -> PERSIST

Each benchmark observation binds:
TASK_ID | V9_PLAN_DIGEST | PROVENANCE_HEAD | RELEASE_STATUS | REGRESSION_STATUS | QUALITY_SCORE | WAVE_COUNT | CONFLICT_COUNT | BLOCKED_COUNT | SPECIALIST_RESULTS.

Adaptive rules

portable.agency_adaptive_planning.recommend_plan consumes prior observations before the next execution.

• Fewer than the minimum samples preserves the requested posture.
• Recent release or artifact-regression failures tighten mutation mode.
• Frequent resource conflicts or blocked work caps supporting specialists.
• Sustained high quality with low conflict may permit one additional support specialist, subject to the caller's requested ceiling.
• Recommendations never grant permissions, alter protected paths, bypass v9 scheduling, or make an unverified promotion decision.

Benchmark persistence

BenchmarkHistory.to_jsonl() / from_jsonl() provide a portable history format. The caller owns durable storage policy; corrupted history must fail closed rather than being silently ignored.

Provenance and regression

The v9 execution-plan digest remains the identity anchor for a run. The provenance head captured at benchmark time provides traceability to the decision chain. Artifact regression status is an explicit learning signal; a failed regression is never treated as a successful benchmark outcome.

Host boundary

The runtime plans and learns; the host executes tools, controls credentials and permissions, enforces sandboxing, owns concurrency/side effects, and decides where benchmark history is persisted. v10 must not autonomously expand authority.

Verification

At minimum run:
python -m py_compile portable/agency_*.py portable/ai_coding_agency_bridge.py
python -m unittest tests/test_agency_adaptive_planning.py tests/test_ai_coding_agency_bridge.py -v
python -m unittest discover -s tests -p 'test_agency_*.py' -v